Internal study · 2026-10-27 · kept
Shutdown wakeups.
All gates passed.
V13 primary paired full-command time changed -94.65% versus immutable V11. The primary and all 23 controls passed the frozen rule; V13 was kept.
V13 is retained as the accepted version; V11 remains the immutable comparison baseline. Interrupt worker heartbeat and leader ticker waits at the existing final stop points. The running heartbeat/Tick cadence, operators, wire format and both timer boundaries remain unchanged.
- -94.65%
- paired full-command change
small-pipeline primary versus V11 - 2,400 / 2,400
- calls with exact complete outputs
- 23 / 23
- controls passed
every control was required
The complete-command primary.
The small-pipeline full-command median was 1.032 seconds for V11 and 0.039 seconds for V13. The paired estimate is -94.65%, with adjusted interval [-96.75%, -91.21%] and exact sign-swap p = 1.907349e-6. The paired estimate is not the ratio of those two medians.
Full-command time includes process and worker startup, execution, shutdown, CSV export, reporting and exit. Native wall retains the existing runLead boundary; spawning, worker reap and later export/reporting remain outside it. Neither timer boundary changed.
All 24 endpoints.
The full-command primary requires an adjusted upper bound ≤ −20% and p < 0.05 / 24. All 11 other full-command cases and all 12 native-wall cases require an adjusted upper bound ≤ +5%. 11 / 11 full-command controls and 12 / 12 native-wall controls passed.
The flagged rows were V12’s three native-wall misses. They are measured anew here against V11; the rejected V12 result is unchanged. See their interval graphics on the progress page →
| Workflow | Full command | Command gate | Native wall | Wall gate |
|---|---|---|---|---|
| Count by product + join · 250k | -86.41%[-89.78%, -81.93%] | Pass | -41.43%[-46.97%, -35.30%] | Pass |
| Count by product + join · 1MV12 native-wall miss · V13 pass | -60.24%[-66.46%, -52.86%] | Pass | -22.89%[-29.00%, -16.25%] | Pass |
| Filter + aggregate + join · 250k | -78.58%[-81.53%, -75.15%] | Pass | -19.27%[-24.96%, -13.14%] | Pass |
| Filter + aggregate + join · 1M | -63.07%[-69.31%, -55.56%] | Pass | -13.30%[-18.42%, -7.86%] | Pass |
| Small pipelineFull-command primary | -94.65%[-96.75%, -91.21%] | Pass | -93.90%[-95.45%, -91.81%] | Pass |
| Duplicate join keys · 64 | -95.03%[-97.14%, -91.37%] | Pass | -94.26%[-95.43%, -92.79%] | Pass |
| Empty right input · 64 | -92.36%[-96.21%, -84.58%] | Pass | -94.35%[-95.48%, -92.93%] | Pass |
| Empty left input · 64 | -93.58%[-97.06%, -85.97%] | Pass | -93.86%[-95.24%, -92.08%] | Pass |
| Pipeline + sort · 1M | -63.67%[-68.71%, -57.83%] | Pass | -11.82%[-15.25%, -8.25%] | Pass |
| Inferred join keys · 1MV12 native-wall miss · V13 pass | -58.93%[-66.01%, -50.37%] | Pass | -23.26%[-29.42%, -16.56%] | Pass |
| Count + join default control · 1MV12 native-wall miss · V13 pass | -59.97%[-66.22%, -52.56%] | Pass | -21.13%[-29.05%, -12.33%] | Pass |
| Sum + sort · 1M | -93.97%[-94.85%, -92.94%] | Pass | -75.65%[-81.32%, -68.25%] | Pass |
Every predeclared gate passed in this study.
Fixed protocol and practical scope.
All 2,400 calls ran once: 12 cases, 20 paired blocks, 5 matched repetitions per version and block. Those repetitions form one geometric mean, leaving 20 inference units. Intervals are paired-t approximations on block log-latency ratios, with 19 degrees of freedom and Bonferroni alpha 0.05 / 24 across the full 24-endpoint family. The p-values use exact two-sided enumeration of all 220 block sign swaps. There were 0 within-study retries and 0 exclusions.
Apple M5 Max, 18 cores, macOS. Three requested loopback workers, four threads each, on one shared Mac. All 1925 contended calls and 78 worker-admission shortfalls remain included. 0 recorded operating anomalies are reported, rather than silently removed.
Every call matched complete old CSV output and printed/native-JSON/journal checksums. Existing native suite: 963 passed cases and 362601 assertions; Cast subset: 16 passed cases and 813 assertions. Goldens retain the same seven known transcript mismatches (634/641) and 14 database skips. Correctness only. Same7golden failures; worker deaths hit Map tasks6and10;9spills/0reloads; two leader restarts loseall3originalworkers and recomputeoldDone; mixed actualDonecoverage includes bothversions. No fresh timing inference.
V12 remains a separate rejected experiment. Its user-paused trial (655 complete calls and one interrupted invocation) and subsequent unpooled 2400-call trial are unchanged. None of those observations enters V13 statistics.
Input identity reads warm the OS cache symmetrically before each call. This is an internal shared-Mac lifecycle comparison against immutable V11, with unchanged operators and both clock boundaries. It establishes no direct Photon, new Rust/Zig/Swift/Go, quiet-host, multi-host or market result.
Source and evidence identifiers
- accepted V11 baseline source commit
- 7d94e68c043c5111f52a624bda8e347b9a9da34b
- V13 candidate source commit
- 97d2ead89b8cbae1a46effed7738de9a106077b4
- final accepted source commit
- 97d2ead89b8cbae1a46effed7738de9a106077b4
- raw backup commit
- 90d501cf03f3e4268f97e7d8b58e363ecfbb9847
- analysis sha256
- 5e8bde7c483308d118b3e8b1b75d1fd45111863119252afd76fa57ada461cbe6
- protocol sha256
- 958213b04e1114d07badfe642568dbdbe9081830c5c945a90eca31a6189f7a7d
- records sha256
- f9b45ea0411f6ca60320f03e0e8760cb95522aee9440c81d859418ce26af4f38
- review sha256
- e21a8a98cf5c0e0fbced9289da1ea9b0a0ac749d4930d6c99d952bd186083777
- statistics sha256
- 1924be22be675072879903987ee62a969a720abd2bf89aac2067aa641303c018
- raw sha256
- 00124fc82c2b5ce463791f0f467de93d995ffe40151d035a074be405864016cc
- state sha256
- 704b08b03c23b2e23d3b3a60e577e9ea0f7b808cbd8b39d4034da0ca7924ca46
- decision sha256
- d9440caeea244772b7bb39eef0f3d8b8cdb3451e1ce05892a1346aa9c0e1774d
- identity sha256
- d09d132422e5f1734e4526b62ccb31e0da35d5047e1c761310d5acf9bd0050aa
- correctness sha256
- f2b6d12e17e0e3a9d8c87b74d8922529565c6bb0efd2639b2789045dec215895
V11, rejected V12 and the historical language tables retain their original measurements. No V13 multiplier is applied to those results. Summary data and hashes are public; engine code, binaries and raw operational logs remain private.